Mammographic Density Change and Risk of Breast Cancer

Shadi Azam; Mikael Eriksson; Arvid Sjölander; Roxanna Hellgren; Marike Gabrielson; Kamila Czene; Per Hall

Disclosures

J Natl Cancer Inst. 2020;112(4):391-399. 

In This Article

Methods

Study Population

KARMA is a population-based prospective screening cohort.[16] All women invited for screening from January 2011 to March 2013 at four mammography units in the national mammography screening program in Sweden were invited to participate in the study. Women with a baseline mammogram (n = 70 874) were included in this study. Informed consent was given for a continuous collection of mammograms. Reasons for exclusion are given in Figure 1, and the final analyses included 43 810 women aged 30–79 years. All participants signed an informed consent, and the ethical review board at Karolinska Institute approved the study.

Figure 1.

Reasons for exclusion of participants in the Karolinska Mammography Project for Risk Prediction of Breast Cancer (KARMA) cohort. BMI = body mass index.

MD Measurement

Processed mammograms from the mediolateral oblique view of left and right breasts were collected from full-field digital mammography systems. For women with BC, we used the mammograms from the contralateral breast (ie, the breast not having a tumor) and for women without BC, we randomly selected mammograms from either left or right breast and followed MD of the same breast to the end of follow-up. Dense area (cm2) and percent MD were measured using the STRATUS method.[15] We chose to present the results mainly using MD area because, in contrast to percent MD, it is less influenced by body mass index (BMI).[17] Nevertheless, for compatibility with other studies, we also presented results using percent MD. We chose to categorize women according to baseline MD area and percent MD in quartile; however, we combined the two highest categories due to few women in the highest quartile. An alternative would have been to use BI-RADS categories A–D,[18] but the distribution of women according to that score makes subgroup analyses unstable.

STRATUS is a fully automated tool developed to analyze digital and analogue images using an algorithm that measures density on all types of images regardless of vendor. STRATUS uses threshold techniques and assesses image features from the mammograms and estimates MD using machine learning. STRATUS measures the MD density area and the breast area and calculates the percent density from these measures.[15] In using the repeated mammograms from the same woman, it is important to take technical differences between mammograms into consideration. Supplementary Figure 1 (available online) illustrates how the breast tissue from the same woman is presented in a mammogram at two points in time. Frame A (Supplementary Figure 1, available online) shows that the same amount of tissue is not found in two mammograms of the same breast taken within minutes, which gives a false difference when comparing density. To minimize the effect of this artifact, mammograms were aligned before MD measurements (Frame B). The concept of alignment and the method has been described previously.[15] In the current study, mammograms from the same woman were aligned before density measures were performed.

Supplementary Figure 1.

Alignment procedure of images. Two mammograms of the same breast were taken 2 minutes apart by the same radiographer. In Frame A the mammograms were superimposed to show the difference in breast placement in the mammography machine. In Frame B, the two images were digitally aligned to the image showing the smallest breast size (outlined with red in Frame A) prior to density measurement. The alignment tool is fully automated and used the National Institute of Health software ImageJ to read pairs of Full Field Digital Mammograms and align the images. First each breast area is marked using a threshold method. The breast area binary masks are used to guide the superimposition of the breast areas on top of each other in layers. The breast area masks are moved towards each other to the optimal position where their pixel intensities show minimal difference in least square means. The positioning technique is therefore not sensitive to aligning images with differences in pixel intensities such as seen in raw, processed or analogue mammograms.

Covariates

At baseline, participants who completed an extensive web-based questionnaire covering established risk factors associated with BC mammograms taken within ±3 months from the date of answering the questionnaires were considered baseline mammograms. Risk factors were categorized as: age at baseline (<50, 50–60, >60 years), BMI (<20, 20–24.9, 25–29.9, ≥30 kg/m2), smoking status (never, former, current), alcohol consumption (0, 0.1–10, >10 g/d), physical activity (<40, 40–44.9, 45.0–49.9, ≥50 metabolic equivalent of task-h/d), age at first birth (<24.9, 25–29.9, ≥30 years), number of children (0, 1–2, ≥3), breast feeding duration (<6, 6–12, >12 months), time since last birth (<10, ≥10 years), age at menarche (<13, ≥13 years), oral contraceptive use (never, ever), menopausal hormone therapy (MHT) status (never, former, current), family history of BC (yes, no), MD area (<9, 9–20, >20 cm2), and percent MD (<5, 5–25, >25%) at baseline. Women reporting no natural menstruation over the past 12 months before study entry or no menstruation due to oophorectomy were considered postmenopausal. Women with missing information on menstruation status or having no menstruation due to gynecological surgeries other than oophorectomy were considered premenopausal if they were age 50 years or younger and postmenopausal if older than 50 years.

Statistical Analyses

Cox proportional hazards regression was used, with age as the underlying time scale, to estimate the association between established risk factors for BC and BC risk. Hazard ratios (HRs) and 95% confidence intervals (CIs) were reported. The proportional-hazards assumption was tested using the Schoenfeld residual test, and no major model violation was observed. All associations were adjusted for age and BMI at baseline.

To investigate the association between area MDC and risk of BC, we carried out analyses in two steps. First, for each woman and consecutive examinations, we calculated the relative area MDC per year between each of the consecutive examinations. Specifically, for two consecutive examinations at time points (ie, ages) t1 and t2, with measured areas MDC, MD1, and MD2, respectively, we defined the relative MDC per year as (MD2 − MD1)/MD1/(t2 − t1). This relative (to MD1) measure reflects the notion that a decrease with, for example, 10 cm2 from MD1 = 50 cm2 is biologically more "dramatic" than a decrease with 10 cm2 from MD1 = 100 cm2. Relative area MDC was categorized as decrease (>10% decrease per year), stable (no change), and increase (>10% increase per year) in agreement with previous literature.[19,20] Second, we used Cox proportional hazard regression to estimate the association of BC with relative area MDC, treating relative MDC as a time-varying exposure in the model. The proportional-hazards assumption was tested using the Schoenfeld residual test, and no major model violation was observed. The associations were adjusted for age, BMI, and dense area (cm2) at baseline, assuming only main effects of (eg, no interactions between) these covariates and relative MDC. Finally, we repeated the analyses, allowing for interactions between relative MDC and baseline MD, to study how these jointly influence the BC risk, and a global test was used to determine the presence of interactions. The same analyses were performed using percent density.

Finally, the Pearson correlation coefficient was used to assess the association between baseline MD area and baseline percent MD with baseline BMI. All statistical analyses were performed with R version 3.4.1. P values, obtained from two-sided Wald/maximum likelihood ratio tests, of less than .05 were considered statistically significant.

Comments

3090D553-9492-4563-8681-AD288FA52ACE

processing....